Scaling Misuse Detection with Self-Assessment
safeguards.anthropic.com
2 threads
So (if my understanding is right) Claude makes and stores/sends a report after every (suspicious, per anthropic) message, and enough reports get flagged?
I've never been more interested in running my LLM locally. Thanks Dario.
Misuse as defined by Anthropic.
Because, of course. Why'd they be the nice guys?
I've had so many conversations trying to use (for example) disclosed CVEs to gain access over my devices, and the lovely guardrails refuse to answer, thinking I'm targeting someone else.
And I suspect even when you later clarify your position in the chat and "convince" Claude, the report's probably gone and already flagged your account.